mm/mprotect.c

   1 /*
   2  *  mm/mprotect.c
   3  *
   4  *  (C) Copyright 1994 Linus Torvalds
   5  *  (C) Copyright 2002 Christoph Hellwig
   6  *
   7  *  Address space accounting code       <alan@lxorguk.ukuu.org.uk>
   8  *  (C) Copyright 2002 Red Hat Inc, All Rights Reserved
   9  */
  10
  11 #include <linux/mm.h>
  12 #include <linux/hugetlb.h>
  13 #include <linux/shm.h>
  14 #include <linux/mman.h>
  15 #include <linux/fs.h>
  16 #include <linux/highmem.h>
  17 #include <linux/security.h>
  18 #include <linux/mempolicy.h>
  19 #include <linux/personality.h>
  20 #include <linux/syscalls.h>
  21 #include <linux/swap.h>
  22 #include <linux/swapops.h>
  23 #include <linux/mmu_notifier.h>
  24 #include <linux/migrate.h>
  25 #include <linux/perf_event.h>
  26 #include <linux/pkeys.h>
  27 #include <linux/ksm.h>
  28 #include <linux/uaccess.h>
  29 #include <asm/pgtable.h>
  30 #include <asm/cacheflush.h>
  31 #include <asm/mmu_context.h>
  32 #include <asm/tlbflush.h>
  33
  34 #include "internal.h"
  35
  36 static unsigned long change_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
  37                 unsigned long addr, unsigned long end, pgprot_t newprot,
  38                 int dirty_accountable, int prot_numa)
  39 {
  40         struct mm_struct *mm = vma->vm_mm;
  41         pte_t *pte, oldpte;
  42         spinlock_t *ptl;
  43         unsigned long pages = 0;
  44         int target_node = NUMA_NO_NODE;
  45
  46         /*
  47          * Can be called with only the mmap_sem for reading by
  48          * prot_numa so we must check the pmd isn't constantly
  49          * changing from under us from pmd_none to pmd_trans_huge
  50          * and/or the other way around.
  51          */
  52         if (pmd_trans_unstable(pmd))
  53                 return 0;
  54
  55         /*
  56          * The pmd points to a regular pte so the pmd can't change
  57          * from under us even if the mmap_sem is only hold for
  58          * reading.
  59          */
  60         pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
  61
  62         /* Get target node for single threaded private VMAs */
  63         if (prot_numa && !(vma->vm_flags & VM_SHARED) &&
  64             atomic_read(&vma->vm_mm->mm_users) == 1)
  65                 target_node = numa_node_id();
  66
  67         flush_tlb_batched_pending(vma->vm_mm);
  68         arch_enter_lazy_mmu_mode();
  69         do {
  70                 oldpte = *pte;
  71                 if (pte_present(oldpte)) {
  72                         pte_t ptent;
  73                         bool preserve_write = prot_numa && pte_write(oldpte);
  74
  75                         /*
  76                          * Avoid trapping faults against the zero or KSM
  77                          * pages. See similar comment in change_huge_pmd.
  78                          */
  79                         if (prot_numa) {
  80                                 struct page *page;
  81
  82                                 page = vm_normal_page(vma, addr, oldpte);
  83                                 if (!page || PageKsm(page))
  84                                         continue;
  85
  86                                 /* Avoid TLB flush if possible */
  87                                 if (pte_protnone(oldpte))
  88                                         continue;
  89
  90                                 /*
  91                                  * Don't mess with PTEs if page is already on the node
  92                                  * a single-threaded process is running on.
  93                                  */
  94                                 if (target_node == page_to_nid(page))
  95                                         continue;
  96                         }
  97
  98                         ptent = ptep_modify_prot_start(mm, addr, pte);
  99                         ptent = pte_modify(ptent, newprot);
 100                         if (preserve_write)
 101                                 ptent = pte_mk_savedwrite(ptent);
 102
 103                         /* Avoid taking write faults for known dirty pages */
 104                         if (dirty_accountable && pte_dirty(ptent) &&
 105                                         (pte_soft_dirty(ptent) ||
 106                                          !(vma->vm_flags & VM_SOFTDIRTY))) {
 107                                 ptent = pte_mkwrite(ptent);
 108                         }
 109                         ptep_modify_prot_commit(mm, addr, pte, ptent);
 110                         pages++;
 111                 } else if (IS_ENABLED(CONFIG_MIGRATION)) {
 112                         swp_entry_t entry = pte_to_swp_entry(oldpte);
 113
 114                         if (is_write_migration_entry(entry)) {
 115                                 pte_t newpte;
 116                                 /*
 117                                  * A protection check is difficult so
 118                                  * just be safe and disable write
 119                                  */
 120                                 make_migration_entry_read(&entry);
 121                                 newpte = swp_entry_to_pte(entry);
 122                                 if (pte_swp_soft_dirty(oldpte))
 123                                         newpte = pte_swp_mksoft_dirty(newpte);
 124                                 set_pte_at(mm, addr, pte, newpte);
 125
 126                                 pages++;
 127                         }
 128
 129                         if (is_write_device_private_entry(entry)) {
 130                                 pte_t newpte;
 131
 132                                 /*
 133                                  * We do not preserve soft-dirtiness. See
 134                                  * copy_one_pte() for explanation.
 135                                  */
 136                                 make_device_private_entry_read(&entry);
 137                                 newpte = swp_entry_to_pte(entry);
 138                                 set_pte_at(mm, addr, pte, newpte);
 139
 140                                 pages++;
 141                         }
 142                 }
 143         } while (pte++, addr += PAGE_SIZE, addr != end);
 144         arch_leave_lazy_mmu_mode();
 145         pte_unmap_unlock(pte - 1, ptl);
 146
 147         return pages;
 148 }
 149
 150 static inline unsigned long change_pmd_range(struct vm_area_struct *vma,
 151                 pud_t *pud, unsigned long addr, unsigned long end,
 152                 pgprot_t newprot, int dirty_accountable, int prot_numa)
 153 {
 154         pmd_t *pmd;
 155         struct mm_struct *mm = vma->vm_mm;
 156         unsigned long next;
 157         unsigned long pages = 0;
 158         unsigned long nr_huge_updates = 0;
 159         unsigned long mni_start = 0;
 160
 161         pmd = pmd_offset(pud, addr);
 162         do {
 163                 unsigned long this_pages;
 164
 165                 next = pmd_addr_end(addr, end);
 166                 if (!is_swap_pmd(*pmd) && !pmd_trans_huge(*pmd) && !pmd_devmap(*pmd)
 167                                 && pmd_none_or_clear_bad(pmd))
 168                         continue;
 169
 170                 /* invoke the mmu notifier if the pmd is populated */
 171                 if (!mni_start) {
 172                         mni_start = addr;
 173                         mmu_notifier_invalidate_range_start(mm, mni_start, end);
 174                 }
 175
 176                 if (is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) || pmd_devmap(*pmd)) {
 177                         if (next - addr != HPAGE_PMD_SIZE) {
 178                                 __split_huge_pmd(vma, pmd, addr, false, NULL);
 179                         } else {
 180                                 int nr_ptes = change_huge_pmd(vma, pmd, addr,
 181                                                 newprot, prot_numa);
 182
 183                                 if (nr_ptes) {
 184                                         if (nr_ptes == HPAGE_PMD_NR) {
 185                                                 pages += HPAGE_PMD_NR;
 186                                                 nr_huge_updates++;
 187                                         }
 188
 189                                         /* huge pmd was handled */
 190                                         continue;
 191                                 }
 192                         }
 193                         /* fall through, the trans huge pmd just split */
 194                 }
 195                 this_pages = change_pte_range(vma, pmd, addr, next, newprot,
 196                                  dirty_accountable, prot_numa);
 197                 pages += this_pages;
 198         } while (pmd++, addr = next, addr != end);
 199
 200         if (mni_start)
 201                 mmu_notifier_invalidate_range_end(mm, mni_start, end);
 202
 203         if (nr_huge_updates)
 204                 count_vm_numa_events(NUMA_HUGE_PTE_UPDATES, nr_huge_updates);
 205         return pages;
 206 }
 207
 208 static inline unsigned long change_pud_range(struct vm_area_struct *vma,
 209                 p4d_t *p4d, unsigned long addr, unsigned long end,
 210                 pgprot_t newprot, int dirty_accountable, int prot_numa)
 211 {
 212         pud_t *pud;
 213         unsigned long next;
 214         unsigned long pages = 0;
 215
 216         pud = pud_offset(p4d, addr);
 217         do {
 218                 next = pud_addr_end(addr, end);
 219                 if (pud_none_or_clear_bad(pud))
 220                         continue;
 221                 pages += change_pmd_range(vma, pud, addr, next, newprot,
 222                                  dirty_accountable, prot_numa);
 223         } while (pud++, addr = next, addr != end);
 224
 225         return pages;
 226 }
 227
 228 static inline unsigned long change_p4d_range(struct vm_area_struct *vma,
 229                 pgd_t *pgd, unsigned long addr, unsigned long end,
 230                 pgprot_t newprot, int dirty_accountable, int prot_numa)
 231 {
 232         p4d_t *p4d;
 233         unsigned long next;
 234         unsigned long pages = 0;
 235
 236         p4d = p4d_offset(pgd, addr);
 237         do {
 238                 next = p4d_addr_end(addr, end);
 239                 if (p4d_none_or_clear_bad(p4d))
 240                         continue;
 241                 pages += change_pud_range(vma, p4d, addr, next, newprot,
 242                                  dirty_accountable, prot_numa);
 243         } while (p4d++, addr = next, addr != end);
 244
 245         return pages;
 246 }
 247
 248 static unsigned long change_protection_range(struct vm_area_struct *vma,
 249                 unsigned long addr, unsigned long end, pgprot_t newprot,
 250                 int dirty_accountable, int prot_numa)
 251 {
 252         struct mm_struct *mm = vma->vm_mm;
 253         pgd_t *pgd;
 254         unsigned long next;
 255         unsigned long start = addr;
 256         unsigned long pages = 0;
 257
 258         BUG_ON(addr >= end);
 259         pgd = pgd_offset(mm, addr);
 260         flush_cache_range(vma, addr, end);
 261         inc_tlb_flush_pending(mm);
 262         do {
 263                 next = pgd_addr_end(addr, end);
 264                 if (pgd_none_or_clear_bad(pgd))
 265                         continue;
 266                 pages += change_p4d_range(vma, pgd, addr, next, newprot,
 267                                  dirty_accountable, prot_numa);
 268         } while (pgd++, addr = next, addr != end);
 269
 270         /* Only flush the TLB if we actually modified any entries: */
 271         if (pages)
 272                 flush_tlb_range(vma, start, end);
 273         dec_tlb_flush_pending(mm);
 274
 275         return pages;
 276 }
 277
 278 unsigned long change_protection(struct vm_area_struct *vma, unsigned long start,
 279                        unsigned long end, pgprot_t newprot,
 280                        int dirty_accountable, int prot_numa)
 281 {
 282         unsigned long pages;
 283
 284         if (is_vm_hugetlb_page(vma))
 285                 pages = hugetlb_change_protection(vma, start, end, newprot);
 286         else
 287                 pages = change_protection_range(vma, start, end, newprot, dirty_accountable, prot_numa);
 288
 289         return pages;
 290 }
 291
 292 int
 293 mprotect_fixup(struct vm_area_struct *vma, struct vm_area_struct **pprev,
 294         unsigned long start, unsigned long end, unsigned long newflags)
 295 {
 296         struct mm_struct *mm = vma->vm_mm;
 297         unsigned long oldflags = vma->vm_flags;
 298         long nrpages = (end - start) >> PAGE_SHIFT;
 299         unsigned long charged = 0;
 300         pgoff_t pgoff;
 301         int error;
 302         int dirty_accountable = 0;
 303
 304         if (newflags == oldflags) {
 305                 *pprev = vma;
 306                 return 0;
 307         }
 308
 309         /*
 310          * If we make a private mapping writable we increase our commit;
 311          * but (without finer accounting) cannot reduce our commit if we
 312          * make it unwritable again. hugetlb mapping were accounted for
 313          * even if read-only so there is no need to account for them here
 314          */
 315         if (newflags & VM_WRITE) {
 316                 /* Check space limits when area turns into data. */
 317                 if (!may_expand_vm(mm, newflags, nrpages) &&
 318                                 may_expand_vm(mm, oldflags, nrpages))
 319                         return -ENOMEM;
 320                 if (!(oldflags & (VM_ACCOUNT|VM_WRITE|VM_HUGETLB|
 321                                                 VM_SHARED|VM_NORESERVE))) {
 322                         charged = nrpages;
 323                         if (security_vm_enough_memory_mm(mm, charged))
 324                                 return -ENOMEM;
 325                         newflags |= VM_ACCOUNT;
 326                 }
 327         }
 328
 329         /*
 330          * First try to merge with previous and/or next vma.
 331          */
 332         pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
 333         *pprev = vma_merge(mm, *pprev, start, end, newflags,
 334                            vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma),
 335                            vma->vm_userfaultfd_ctx);
 336         if (*pprev) {
 337                 vma = *pprev;
 338                 VM_WARN_ON((vma->vm_flags ^ newflags) & ~VM_SOFTDIRTY);
 339                 goto success;
 340         }
 341
 342         *pprev = vma;
 343
 344         if (start != vma->vm_start) {
 345                 error = split_vma(mm, vma, start, 1);
 346                 if (error)
 347                         goto fail;
 348         }
 349
 350         if (end != vma->vm_end) {
 351                 error = split_vma(mm, vma, end, 0);
 352                 if (error)
 353                         goto fail;
 354         }
 355
 356 success:
 357         /*
 358          * vm_flags and vm_page_prot are protected by the mmap_sem
 359          * held in write mode.
 360          */
 361         vma->vm_flags = newflags;
 362         dirty_accountable = vma_wants_writenotify(vma, vma->vm_page_prot);
 363         vma_set_page_prot(vma);
 364
 365         change_protection(vma, start, end, vma->vm_page_prot,
 366                           dirty_accountable, 0);
 367
 368         /*
 369          * Private VM_LOCKED VMA becoming writable: trigger COW to avoid major
 370          * fault on access.
 371          */
 372         if ((oldflags & (VM_WRITE | VM_SHARED | VM_LOCKED)) == VM_LOCKED &&
 373                         (newflags & VM_WRITE)) {
 374                 populate_vma_page_range(vma, start, end, NULL);
 375         }
 376
 377         vm_stat_account(mm, oldflags, -nrpages);
 378         vm_stat_account(mm, newflags, nrpages);
 379         perf_event_mmap(vma);
 380         return 0;
 381
 382 fail:
 383         vm_unacct_memory(charged);
 384         return error;
 385 }
 386
 387 /*
 388  * pkey==-1 when doing a legacy mprotect()
 389  */
 390 static int do_mprotect_pkey(unsigned long start, size_t len,
 391                 unsigned long prot, int pkey)
 392 {
 393         unsigned long nstart, end, tmp, reqprot;
 394         struct vm_area_struct *vma, *prev;
 395         int error = -EINVAL;
 396         const int grows = prot & (PROT_GROWSDOWN|PROT_GROWSUP);
 397         const bool rier = (current->personality & READ_IMPLIES_EXEC) &&
 398                                 (prot & PROT_READ);
 399
 400         prot &= ~(PROT_GROWSDOWN|PROT_GROWSUP);
 401         if (grows == (PROT_GROWSDOWN|PROT_GROWSUP)) /* can't be both */
 402                 return -EINVAL;
 403
 404         if (start & ~PAGE_MASK)
 405                 return -EINVAL;
 406         if (!len)
 407                 return 0;
 408         len = PAGE_ALIGN(len);
 409         end = start + len;
 410         if (end <= start)
 411                 return -ENOMEM;
 412         if (!arch_validate_prot(prot))
 413                 return -EINVAL;
 414
 415         reqprot = prot;
 416
 417         if (down_write_killable(&current->mm->mmap_sem))
 418                 return -EINTR;
 419
 420         /*
 421          * If userspace did not allocate the pkey, do not let
 422          * them use it here.
 423          */
 424         error = -EINVAL;
 425         if ((pkey != -1) && !mm_pkey_is_allocated(current->mm, pkey))
 426                 goto out;
 427
 428         vma = find_vma(current->mm, start);
 429         error = -ENOMEM;
 430         if (!vma)
 431                 goto out;
 432         prev = vma->vm_prev;
 433         if (unlikely(grows & PROT_GROWSDOWN)) {
 434                 if (vma->vm_start >= end)
 435                         goto out;
 436                 start = vma->vm_start;
 437                 error = -EINVAL;
 438                 if (!(vma->vm_flags & VM_GROWSDOWN))
 439                         goto out;
 440         } else {
 441                 if (vma->vm_start > start)
 442                         goto out;
 443                 if (unlikely(grows & PROT_GROWSUP)) {
 444                         end = vma->vm_end;
 445                         error = -EINVAL;
 446                         if (!(vma->vm_flags & VM_GROWSUP))
 447                                 goto out;
 448                 }
 449         }
 450         if (start > vma->vm_start)
 451                 prev = vma;
 452
 453         for (nstart = start ; ; ) {
 454                 unsigned long mask_off_old_flags;
 455                 unsigned long newflags;
 456                 int new_vma_pkey;
 457
 458                 /* Here we know that vma->vm_start <= nstart < vma->vm_end. */
 459
 460                 /* Does the application expect PROT_READ to imply PROT_EXEC */
 461                 if (rier && (vma->vm_flags & VM_MAYEXEC))
 462                         prot |= PROT_EXEC;
 463
 464                 /*
 465                  * Each mprotect() call explicitly passes r/w/x permissions.
 466                  * If a permission is not passed to mprotect(), it must be
 467                  * cleared from the VMA.
 468                  */
 469                 mask_off_old_flags = VM_READ | VM_WRITE | VM_EXEC |
 470                                         ARCH_VM_PKEY_FLAGS;
 471
 472                 new_vma_pkey = arch_override_mprotect_pkey(vma, prot, pkey);
 473                 newflags = calc_vm_prot_bits(prot, new_vma_pkey);
 474                 newflags |= (vma->vm_flags & ~mask_off_old_flags);
 475
 476                 /* newflags >> 4 shift VM_MAY% in place of VM_% */
 477                 if ((newflags & ~(newflags >> 4)) & (VM_READ | VM_WRITE | VM_EXEC)) {
 478                         error = -EACCES;
 479                         goto out;
 480                 }
 481
 482                 error = security_file_mprotect(vma, reqprot, prot);
 483                 if (error)
 484                         goto out;
 485
 486                 tmp = vma->vm_end;
 487                 if (tmp > end)
 488                         tmp = end;
 489                 error = mprotect_fixup(vma, &prev, nstart, tmp, newflags);
 490                 if (error)
 491                         goto out;
 492                 nstart = tmp;
 493
 494                 if (nstart < prev->vm_end)
 495                         nstart = prev->vm_end;
 496                 if (nstart >= end)
 497                         goto out;
 498
 499                 vma = prev->vm_next;
 500                 if (!vma || vma->vm_start != nstart) {
 501                         error = -ENOMEM;
 502                         goto out;
 503                 }
 504                 prot = reqprot;
 505         }
 506 out:
 507         up_write(&current->mm->mmap_sem);
 508         return error;
 509 }
 510
 511 SYSCALL_DEFINE3(mprotect, unsigned long, start, size_t, len,
 512                 unsigned long, prot)
 513 {
 514         return do_mprotect_pkey(start, len, prot, -1);
 515 }
 516
 517 #ifdef CONFIG_ARCH_HAS_PKEYS
 518
 519 SYSCALL_DEFINE4(pkey_mprotect, unsigned long, start, size_t, len,
 520                 unsigned long, prot, int, pkey)
 521 {
 522         return do_mprotect_pkey(start, len, prot, pkey);
 523 }
 524
 525 SYSCALL_DEFINE2(pkey_alloc, unsigned long, flags, unsigned long, init_val)
 526 {
 527         int pkey;
 528         int ret;
 529
 530         /* No flags supported yet. */
 531         if (flags)
 532                 return -EINVAL;
 533         /* check for unsupported init values */
 534         if (init_val & ~PKEY_ACCESS_MASK)
 535                 return -EINVAL;
 536
 537         down_write(&current->mm->mmap_sem);
 538         pkey = mm_pkey_alloc(current->mm);
 539
 540         ret = -ENOSPC;
 541         if (pkey == -1)
 542                 goto out;
 543
 544         ret = arch_set_user_pkey_access(current, pkey, init_val);
 545         if (ret) {
 546                 mm_pkey_free(current->mm, pkey);
 547                 goto out;
 548         }
 549         ret = pkey;
 550 out:
 551         up_write(&current->mm->mmap_sem);
 552         return ret;
 553 }
 554
 555 SYSCALL_DEFINE1(pkey_free, int, pkey)
 556 {
 557         int ret;
 558
 559         down_write(&current->mm->mmap_sem);
 560         ret = mm_pkey_free(current->mm, pkey);
 561         up_write(&current->mm->mmap_sem);
 562
 563         /*
 564          * We could provie warnings or errors if any VMA still
 565          * has the pkey set here.
 566          */
 567         return ret;
 568 }
 569
 570 #endif /* CONFIG_ARCH_HAS_PKEYS */